anm-pinellia / dota_yolov2

DOTA_YOLOv2 provides the data convertion code, parameter files while training DOTA using YOLOv2, and the trained model is also provided. So it's convenient for you to use them.

Our code is tested on official darknet@(commit f6d8617) with cuda-8.0 and cudnn-6.0 on Ubuntu 16.04.1 LTS.

Installation

• install darknet
  See Installing Darknet for instructions.
  
• development kit
  The Development kit provides the following functions. You can easily install it following the instructions.
  
  • Load and visulize the data.
  • Evaluate the result.
  • Split and merge the picture and label.
    

Training YOLO on DOTA

• Get the DOTA Dataset
  

• Convert the Label Format
  In DOTA, the annotation format is:

      x1 y1 x2 y2 x3 y3 x4 y4 category difficult
  

  While Darknet wants a .txt file for each image with a line for each ground truth object in the image that looks like:

      category-id x y width height
  

  Where x, y, width, and height are relative to the image's width and height. Here, you can refer to data_transform/YOLO_Transform.py to convert the format.
  

  Note that this code is for the image of size 1024*1024. If not, you should modify it accroding to your image size. For DOTA, you can refer to DOTA_devkit/ImgSplit.py to split the images and labels.

• Modify Cfg for Your Data
  You have to change the cfg/dota.data config file to point to your data:

      classes=15
      train  = /home/yh/dota/dota_data/YOLO/train/train.txt
      valid  = /home/yh/dota/dota_data/YOLO/test/test.txt
      names = data/dota.names
      backup = /home/yh/dota/darknet/dota-backup
  

  You should replace the path here with the path where you put your corresponding file. And the text files like train.txt or test.txt list the image files for training or test. Notice that we use the full path of the image instead of the file name.

• Train the Model
  

      wget https://pjreddie.com/media/files/darknet19_448.conv.23
      sh train-dota.sh 
  

• Evaluate the Results
  You can download the pre-trained model on DOTA from Baidu Drive or Google Drive, and use it to test all the test images.

      sh valid-dota.sh 
  

  Then you will obtain 15 files stored in the results/ subdirectory, and each file contains all the results for a specific category.Each file is in the following format:
  

      imgname score xmin ymin xmax ymax 
  

  If you have split the images before, please first use DOTA_devkit/ResultMerge.py to merge the results.
  
  For DOTA, You can submit your results on the Evaluation Server for evaluation. See the official website of DOTA for details.